Andrej Karpathy taught a neural network to play Pong with 130 lines of Python and zero AI libraries.No PyTorch, no TensorFlow just numpy and a MacBook that ran for 2 days straight.The network saw nothing a human sees. No paddles, no ball, no game. Just 6,400 raw pixels flipping like static, feeding into 1.3 million parameters that started as pure random noise.His training method sounds almost too dumb to work: let the network act randomly, watch what happens, then make the winning moves more lik